pith. sign in

arxiv: 2606.11922 · v1 · pith:L6MCLJRWnew · submitted 2026-06-10 · 💻 cs.SD · cs.AI

Lung-SRAD: Spectral-Aware Regularized Audio DASS with Dual-Axis Patch-Mix Contrastive Learning for Respiratory Sound Classification

Hemansh Shridhar , Miika Toikkanen , June-Woo Kim This is my paper

Pith reviewed 2026-06-27 08:14 UTC · model grok-4.3

classification 💻 cs.SD cs.AI
keywords respiratory sound classificationstate space modelsspectral regularizationcontrastive learningICBHI benchmarkaudio spectrogrampatch-mix augmentation
0
0 comments X

The pith

State space models augmented with spectral regularization and dual-axis contrastive learning raise respiratory sound classification accuracy to 64.48 percent on the ICBHI benchmark.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper replaces attention-based backbones such as the Audio Spectrogram Transformer with Distilled Audio State Space models for respiratory sound classification. Analysis of intermediate representations reveals that these models retain more mid-to-high spatial-frequency content than self-attention layers. To exploit this property the authors add Gaussian-convolution regularization on selected SSM layers and introduce Dual-Axis Patch-Mix contrastive learning designed for SSM audio architectures. On the ICBHI benchmark the resulting model reaches 64.48 percent, five points above the AST baseline. A reader would care because the frequency-preservation difference is offered as the reason the model detects localized abnormal lung-sound patterns more reliably.

Core claim

State Space Models exhibit stronger preservation of mid-to-high spatial-frequency components in their intermediate representations than CLS-token self-attention models. Applying spectral-aware layer regularization via Gaussian convolution to selected layers together with Dual-Axis Patch-Mix contrastive learning produces a model that scores 64.48 percent on the ICBHI respiratory-sound benchmark, outperforming the AST baseline by five percentage points.

What carries the argument

Spectral-aware layer regularization that applies Gaussian convolution to selected SSM layers, combined with Dual-Axis Patch-Mix contrastive learning that operates on both time and frequency patch axes.

If this is right

  • SSM backbones can replace attention layers in audio tasks where local spectral detail matters more than global context.
  • Gaussian convolution regularization can be applied selectively to any SSM layer stack to control frequency response.
  • Dual-axis patch-mix contrastive learning provides a training objective that works directly with the patch structure of SSM audio encoders.
  • The 5-point gain on ICBHI suggests that frequency-aware regularization may transfer to other sound-classification domains that rely on localized cues.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same spectral-response diagnostic could be run on other sequence models to decide when SSMs are preferable to attention.
  • If the frequency-preservation advantage holds, the method could be tested on speech or music datasets that contain fine-grained local events.
  • The regularization strength and choice of which layers receive the Gaussian convolution remain tunable parameters that future work can optimize per dataset.

Load-bearing premise

The stronger preservation of mid-to-high frequencies observed in SSM representations is what enables better detection of localized abnormal patterns in respiratory audio.

What would settle it

An ablation that keeps the SSM backbone but removes the Gaussian spectral regularization and measures whether accuracy falls back to or below the AST level.

Figures

Figures reproduced from arXiv: 2606.11922 by Hemansh Shridhar, June-Woo Kim, Miika Toikkanen.

Figure 1
Figure 1. Figure 1: Spectral filter responses and attention maps of fine-tuned DASS (blue) and AST (red) on the ICBHI dataset. DASS Stage 2 layers (blocks) retain prominent mid-to-high frequency components while AST exhibits dominant low-pass component at all depths. 3. Methodology We introduce Lung-SRAD (Spectral-Aware Regularized Audio DASS for Lung Sounds) to address architectural limitations and spectral biases for RSC, w… view at source ↗
Figure 2
Figure 2. Figure 2: Spectral responses for DASS baseline (blue) and DASS with Gaussian smoothing (red) for Stage 2 Blocks 2–3. Gaussian smoothing attenuates dominant spectral peaks while preserving the harmonic structure of the frequency response. gregates features from the full time–frequency feature map via spatial pooling, allowing the classifier to utilize distributed rep￾resentations rather than a single token. As observ… view at source ↗
Figure 3
Figure 3. Figure 3: Hyperparameter analysis of Patch-Mix CL and spectral-aware regularization, evaluated by ICBHI Score. from task-aligned Patch-Mix designs than the AST-style vari￾ant: frequency-only and time-only mixing improve the Score to 62.22% → 63.18% and 62.22% → 62.83%, respectively, while the proposed Dual-Axis Patch-Mix (Lung-SRAD) achieves the best overall Score of 64.48%, indicating that jointly mixing both axes … view at source ↗
read the original abstract

Recent respiratory sound classification (RSC) studies largely rely on CLS-token driven self-attention architectures such as the Audio Spectrogram Transformer (AST). While effective at modeling global context, recent analyses suggest a low-pass filtering behavior that may reduce sensitivity to localized abnormal patterns. In this work, we investigate State Space Models (SSMs) as an alternative backbone for RSC. Using the Distilled Audio State Space model, we analyze intermediate representations through spectral response curves and observe stronger preservation of mid-to-high spatial-frequency components. Based on these observations, we introduce spectral-aware layer regularization using Gaussian convolution applied to selected layers. We further propose Dual-Axis Patch-Mix contrastive learning tailored to SSM-based audio models for robust representation learning. Experiments on the ICBHI benchmark show that our approach achieves 64.48% score, outperforming the AST baseline by 5%. Code is available at https://github.com/RSC-Toolkit/Lung-SRAD.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Lung-SRAD, which replaces the Audio Spectrogram Transformer (AST) backbone with a Distilled Audio State Space Model (DASS) for respiratory sound classification. It observes stronger mid-to-high spatial-frequency preservation in SSM intermediate representations via spectral response curves, introduces spectral-aware Gaussian convolution regularization on selected layers, and adds Dual-Axis Patch-Mix contrastive learning. On the ICBHI benchmark the method reports a 64.48% score, a 5% absolute improvement over the AST baseline.

Significance. If the reported gain is reproducible and the frequency-preservation mechanism is shown to be causal, the work would supply concrete evidence that SSM inductive biases can outperform global self-attention for tasks that require sensitivity to localized high-frequency events in spectrograms. The provision of public code at https://github.com/RSC-Toolkit/Lung-SRAD is a positive factor for reproducibility.

major comments (2)
  1. [Experimental results] Experimental results section: the central performance claim (64.48% ICBHI score, +5% over AST) is presented without ablation tables that isolate the spectral-aware Gaussian regularization from the Dual-Axis Patch-Mix contrastive objective and the SSM backbone. Because the paper's narrative attributes the gain specifically to mid-to-high frequency retention, the absence of these controls leaves the causal link untested.
  2. [Spectral analysis] Spectral analysis subsection: the spectral response curves are used to motivate the regularization, yet no quantitative correlation (e.g., layer-wise frequency-retention metric versus per-class accuracy) is reported to link the observed curves directly to the final classification improvement.
minor comments (2)
  1. [Abstract] The abstract states the ICBHI 'score' without reminding readers that the official metric is (sensitivity + specificity)/2; a parenthetical clarification would improve readability.
  2. [Results] No error bars or number of random seeds are mentioned for the 64.48% figure; adding these in the results table would strengthen the empirical claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on manuscript arXiv:2606.11922. We address the major comments point by point below and will incorporate the requested analyses to strengthen the causal claims in the revised version.

read point-by-point responses

  1. Referee: [Experimental results] Experimental results section: the central performance claim (64.48% ICBHI score, +5% over AST) is presented without ablation tables that isolate the spectral-aware Gaussian regularization from the Dual-Axis Patch-Mix contrastive objective and the SSM backbone. Because the paper's narrative attributes the gain specifically to mid-to-high frequency retention, the absence of these controls leaves the causal link untested.

    Authors: We agree that the current version lacks explicit ablations isolating each component. In the revision we will add ablation tables in the Experimental results section that separately disable the spectral-aware Gaussian regularization, the Dual-Axis Patch-Mix contrastive objective, and the DASS backbone while keeping the other elements fixed. These tables will report ICBHI scores for each variant, directly testing the contribution of mid-to-high frequency retention to the reported 64.48% score. revision: yes

  2. Referee: [Spectral analysis] Spectral analysis subsection: the spectral response curves are used to motivate the regularization, yet no quantitative correlation (e.g., layer-wise frequency-retention metric versus per-class accuracy) is reported to link the observed curves directly to the final classification improvement.

    Authors: We acknowledge the absence of a quantitative link. In the revised Spectral analysis subsection we will define a layer-wise frequency-retention metric from the spectral response curves and compute its Pearson correlation with per-class accuracy gains. The resulting correlation coefficients and scatter plots will be reported to provide direct evidence connecting the observed frequency preservation to classification performance. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical result from full model training on benchmark

full rationale

The paper's central result is an empirical ICBHI score of 64.48% obtained by training the proposed model (SSM backbone + spectral-aware Gaussian regularization + Dual-Axis Patch-Mix contrastive loss). No equations, fitted parameters, or self-citations are shown that would make this score equivalent to its inputs by construction. The spectral observations motivate the regularization design but do not define the final metric; the performance number remains an independent experimental outcome.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review performed on abstract alone; no free parameters, axioms, or invented entities can be enumerated beyond the high-level modeling choice of SSMs over transformers.

axioms (1)
  • domain assumption State-space models preserve mid-to-high frequency content better than self-attention transformers when processing audio spectrograms.
    Invoked to justify the backbone switch and the subsequent regularization design.

pith-pipeline@v0.9.1-grok · 5714 in / 1213 out tokens · 20553 ms · 2026-06-27T08:14:07.653228+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 3 linked inside Pith

  1. [1]

    Introduction Abnormal lung sounds likecracklesandwheezesare key in- dicators of respiratory disorders such as pneumonia, COPD, and asthma, which account for nearly four million deaths annu- ally [1]. Crackles are associated with diseases affecting the lung parenchyma and manifest as discontinuous acoustic events [2], whereas wheezes are characterized by c...

  2. [2]

    Preliminaries 2.1. Dataset Description We use the ICBHI [22] dataset, comprising 5.5 hours of record- ings and 6,898 breathing cycles with four lung sounds:Normal, Crackle, WheezeandCrackle + Wheeze (Both). We follow the official 60% train and 40% test sets split, with no patient overlap between them, resulting in 4,142 cycles and 2,756 cycles. 2.2. Train...

    Pith/arXiv arXiv 2026

  3. [3]

    Methodology We introduceLung-SRAD(Spectral-Aware Regularized Audio DASS for Lung Sounds) to address architectural limitations and spectral biases for RSC, with theoretical motivation and empir- ical validation for each design choice. 3.1. State Space Models (SSMs) SSMs.SSMs map an input sequencex(t)∈R D to an output y(t)through a latent stateh(t)∈R N : h′...

    arXiv

  4. [4]

    Main Results 4-Class Results.Table 1 summarizes the performance on the ICBHI benchmark under the official 60–40 patient indepen- dent split

    Results 4.1. Main Results 4-Class Results.Table 1 summarizes the performance on the ICBHI benchmark under the official 60–40 patient indepen- dent split. With simple fine-tuning from AudioSet-distilled DASS initialization, the model achieves 61.06% Score. Apply- ing spectral-aware regularization with Gaussian convolution on selected intermediate layers im...

  5. [5]

    We observed that the DASS architecture preserves mid-to-high spatial-frequency components important for capturing local- ized abnormal respiratory events

    Conclusion In this work, we explored SSM as a backbone for RSC. We observed that the DASS architecture preserves mid-to-high spatial-frequency components important for capturing local- ized abnormal respiratory events. Based on this observation, we introduced spectral-aware regularization using Gaussian smoothing on selected intermediate layers, along wit...

  6. [6]

    RS-2025-16066662)

    Acknowledgement This research was supported by the Regional Innovation System & Education(RISE) program through the Jeonbuk RISE Center, funded by the Ministry of Education(MOE) and the Jeonbuk State, Republic of Korea(2026-RISE-13-WKU), and by the Na- tional Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (grant no. RS-2025-16066662)

    2026

  7. [7]

    The authors have verified all technical content and maintain full account- ability for the work

    Generative AI Use Disclosure Generative AI (ChatGPT) was used solely for grammar correc- tion and linguistic polishing of this manuscript. The authors have verified all technical content and maintain full account- ability for the work

  8. [8]

    Assessing the impact of new technologies on man- aging chronic respiratory diseases,

    O. Gra ˜na-Castro, E. Izquierdo, A. Pi ˜nas-Mesa, E. Menasalvas, and T. Chivato-P´erez, “Assessing the impact of new technologies on man- aging chronic respiratory diseases,”Journal of Clinical Medicine, vol. 13, no. 22, p. 6913, 2024

    2024

  9. [9]

    Automated analysis of crackles in patients with interstitial pulmonary fibrosis,

    B. Flietstra, N. Markuzon, A. Vyshedskiy, and R. Murphy, “Automated analysis of crackles in patients with interstitial pulmonary fibrosis,” Pulmonary medicine, vol. 2011, no. 1, p. 590506, 2011

    2011

  10. [10]

    Fundamentals of lung auscultation,

    A. Bohadana, G. Izbicki, and S. S. Kraman, “Fundamentals of lung auscultation,”New England Journal of Medicine, vol. 370, no. 8, pp. 744–751, 2014

    2014

  11. [11]

    Anomaly detection from multivariate time- series with sparse representation,

    N. Takeishi and T. Yairi, “Anomaly detection from multivariate time- series with sparse representation,” in2014 ieee international confer- ence on systems, man, and cybernetics (smc). IEEE, 2014, pp. 2651– 2656

    2014

  12. [12]

    Stethoscope- guided supervised contrastive learning for cross-domain adaptation on respiratory sound classification,

    J.-W. Kim, S. Bae, W.-Y . Cho, B. Lee, and H.-Y . Jung, “Stethoscope- guided supervised contrastive learning for cross-domain adaptation on respiratory sound classification,” inICASSP 2024-2024 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 1431–1435

    2024

  13. [13]

    Adaptive metadata-guided super- vised contrastive learning for domain adaptation on respiratory sound classification,

    J.-W. Kim, M. Toikkanen, A. Jalali, M. Kim, H.-J. Han, H. Kim, W. Shin, H.-Y . Jung, and K. Kim, “Adaptive metadata-guided super- vised contrastive learning for domain adaptation on respiratory sound classification,”IEEE Journal of Biomedical and Health Informatics, 2025

    2025

  14. [14]

    Ast: Audio spectrogram trans- former,

    Y . Gong, Y .-A. Chung, and J. Glass, “Ast: Audio spectrogram trans- former,” inProc. Interspeech 2021, 2021, pp. 571–575

    2021

  15. [15]

    Patch-mix contrastive learning with audio spectrogram transformer on respiratory sound classification,

    S. Bae, J.-W. Kim, W.-Y . Cho, H. Baek, S. Son, B. Lee, C. Ha, K. Tae, S. Kim, and S.-Y . Yun, “Patch-mix contrastive learning with audio spectrogram transformer on respiratory sound classification,” inProc. Interspeech 2023, 2023, pp. 5436–5440

    2023

  16. [16]

    Adver- sarial fine-tuning using generated respiratory sound to address class imbalance,

    J.-W. Kim, C. Yoon, M. Toikkanen, S. Bae, and H.-Y . Jung, “Adver- sarial fine-tuning using generated respiratory sound to address class imbalance,”arXiv preprint arXiv:2311.06480, 2023

    arXiv 2023

  17. [17]

    Repaug- ment: Input-agnostic representation-level augmentation for respiratory sound classification,

    J.-W. Kim, M. Toikkanen, S. Bae, M. Kim, and H.-Y . Jung, “Repaug- ment: Input-agnostic representation-level augmentation for respiratory sound classification,” in2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 2024, pp. 1–6

    2024

  18. [18]

    Tri-mtl: A triple multitask learning approach for respiratory disease diagnosis,

    J.-W. Kim, S. Lee, M. Toikkanen, D. Hwang, and K. Kim, “Tri-mtl: A triple multitask learning approach for respiratory disease diagnosis,” in 2025 47th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 2025, pp. 1–6

    2025

  19. [19]

    Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,

    Y . Wu, K. Chen, T. Zhang, Y . Hui, T. Berg-Kirkpatrick, and S. Dub- nov, “Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  20. [20]

    Bts: Bridging text and sound modalities for metadata-aided respira- tory sound classification,

    J.-W. Kim, M. Toikkanen, Y . Choi, S.-E. Moon, and H.-Y . Jung, “Bts: Bridging text and sound modalities for metadata-aided respira- tory sound classification,” inProc. Interspeech 2024, 2024, pp. 1690– 1694

    2024

  21. [21]

    Improving respiratory sound classifi- cation with architecture-agnostic knowledge distillation from ensem- bles,

    M. Toikkanen and J.-W. Kim, “Improving respiratory sound classifi- cation with architecture-agnostic knowledge distillation from ensem- bles,” inProc. Interspeech 2025, 2025, pp. 1023–1027

    2025

  22. [22]

    Empow- ering multimodal respiratory sound classification with counterfactual adversarial debiasing for out-of-distribution robustness,

    H. Koo, M. Toikkanen, Y . T. Kim, S. Y . Kim, and J.-W. Kim, “Empow- ering multimodal respiratory sound classification with counterfactual adversarial debiasing for out-of-distribution robustness,” inICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 14 967–14 971

    2026

  23. [23]

    Beats: audio pre-training with acoustic tokeniz- ers,

    S. Chen, Y . Wu, C. Wang, S. Liu, D. Tompkins, Z. Chen, W. Che, X. Yu, and F. Wei, “Beats: audio pre-training with acoustic tokeniz- ers,” inProceedings of the 40th International Conference on Machine Learning, 2023, pp. 5178–5193

    2023

  24. [24]

    Patient-aware feature alignment for robust lung sound classification: Cohesion-separation and global alignment losses,

    S. G. Jeong and S. E. Kim, “Patient-aware feature alignment for robust lung sound classification: Cohesion-separation and global alignment losses,” inProc. Interspeech 2025, 2025, pp. 1018–1022

    2025

  25. [25]

    Masked modeling duo: Towards a universal audio pre-training frame- work,

    D. Niizumi, D. Takeuchi, Y . Ohishi, N. Harada, and K. Kashino, “Masked modeling duo: Towards a universal audio pre-training frame- work,”IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 32, pp. 2391–2406, 2024

    2024

  26. [26]

    Anti-oversmoothing in deep vision transformers via the fourier domain analysis: From theory to practice,

    P. Wang, W. Zheng, T. Chen, and Z. Wang, “Anti-oversmoothing in deep vision transformers via the fourier domain analysis: From theory to practice,” inInternational Conference on Learning Representations, 2022. [Online]. Available: https://openreview.net/ forum?id=O476oWmiNNp

    2022

  27. [27]

    Attention sinks: A’catch, tag, release’mechanism for embeddings,

    S. Zhang, M. Khan, and V . Papyan, “Attention sinks: A’catch, tag, release’mechanism for embeddings,”Advances in Neural Information Processing Systems, vol. 38, pp. 83 140–83 181, 2026

    2026

  28. [28]

    Dass: Distilled audio state space models are stronger and more duration-scalable learners,

    S. Bhati, Y . Gong, L. Karlinsky, H. Kuehne, R. Feris, and J. Glass, “Dass: Distilled audio state space models are stronger and more duration-scalable learners,” in2024 IEEE Spoken Language Technol- ogy Workshop (SLT). IEEE, 2024, pp. 1015–1022

    2024

  29. [29]

    A respiratory sound database for the development of automated classification,

    B. M. Rocha, D. Filos, L. Mendes, I. V ogiatzis, E. Perantoni, E. Kaimakamis, P. Natsiavas, A. Oliveira, C. J ´acome, A. Marques et al., “A respiratory sound database for the development of automated classification,” inInternational conference on biomedical and health informatics. Springer, 2017, pp. 33–37

    2017

  30. [30]

    Specaugment: A simple data augmentation method for automatic speech recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “Specaugment: A simple data augmentation method for automatic speech recognition,” inProc. Interspeech 2019, 2019, pp. 2613–2617

    2019

  31. [31]

    Adam: A method for stochastic optimiza- tion,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,”arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  32. [32]

    Mamba: Linear-time sequence modeling with se- lective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with se- lective state spaces,”arXiv preprint arXiv:2312.00752, 2023

    Pith/arXiv arXiv 2023

  33. [33]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, and Y . Liu, “Vmamba: Visual state space model,”Advances in neural in- formation processing systems, vol. 37, pp. 103 031–103 063, 2024

    2024

  34. [34]

    Audio set: An ontology and human-labeled dataset for audio events,

    J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” in2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2017, pp. 776–780

    2017

  35. [35]

    Hts-at: A hierarchical token-semantic audio transformer for sound classification and detection,

    K. Chen, X. Du, B. Zhu, Z. Ma, T. Berg-Kirkpatrick, and S. Dub- nov, “Hts-at: A hierarchical token-semantic audio transformer for sound classification and detection,” inICASSP 2022-2022 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 646–650

    2022

  36. [36]

    Efficient streaming language models with attention sinks,

    G. Xiao, Y . Tian, B. Chen, S. Han, and M. Lewis, “Efficient streaming language models with attention sinks,” inInternational Conference on Learning Representations, vol. 2024, 2024, pp. 21 875–21 895

    2024

  37. [37]

    Ima- genet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Ima- genet: A large-scale hierarchical image database,” in2009 IEEE con- ference on computer vision and pattern recognition. Ieee, 2009, pp. 248–255

    2009

  38. [38]

    Lungadapter: Efficient adapt- ing audio spectrogram transformer for lung sound classification,

    L. Xiao, L. Fang, Y . Yang, and W. Tu, “Lungadapter: Efficient adapt- ing audio spectrogram transformer for lung sound classification,” in Proc. Interspeech 2024, 2024, pp. 4738–4742

    2024